Where Story and Media Meet: Computer Generation of Narrative Discourse. RONFARD, Rémi, SZILAS, Nicolas. Abstract

Transcription

1 Proceedings Chapter Where Story and Media Meet: Computer Generation of Narrative Discourse RONFARD, Rémi, SZILAS, Nicolas Abstract Story generation (including interactive narrative) consists of creating a narrative experience on computer by generating narrative events. It requires building an abstract computational model that can generate a variety of narrative events from a limited set of authored content. These models implement a "story logic", as they formalize the occurrence of an event in the story according to various algorithms. At the same time, these stories aim to be expressed to an audience using digital media, which requires a medium logic. In this contribution, we look at the relation between story logic and medium logic in the production of mediated narrative discourse. Using the terminology of Russian formalists and a metaphor borrowed from cinema production, we introduce three models of increasing complexity. In the first model, the story logic (fabulist) creates a fabula which is performed by the medium logic (director) to a screenplay then to the screen. In the second model, the story logic (screenwriter) generates a sjuzhet composed of narrative discourse acts that are staged by the medium logic (director). In the third model, the [...] Reference RONFARD, Rémi, SZILAS, Nicolas. Where Story and Media Meet: Computer Generation of Narrative Discourse. In: M.A. Finlayson, J.C. Meister and E.G. Bruneau Workshop on Computational Models of Narrative. Dagstuhl, Germany : Schloss Dagstuhl Leibniz-Zentrum fuer Informatik, p Available at: Disclaimer: layout of this document may differ from the published version.

2 Where story and media meet: computer generation of narrative discourse Remi Ronfard 1 and Nicolas Szilas 2 1 Inria, Univ. Grenoble Alpes & CNRS (LJK), Grenoble, France remi.ronfard@inria.fr 2 University of Geneva, Switzerland nicolas.szilas@unige.ch Abstract Story generation (including interactive narrative) consists of creating a narrative experience on computer by generating narrative events. It requires building an abstract computational model that can generate a variety of narrative events from a limited set of authored content. These models implement a story logic, as they formalize the occurrence of an event in the story according to various algorithms. At the same time, these stories aim to be expressed to an audience using digital media, which requires a medium logic. In this contribution, we look at the relation between story logic and medium logic in the production of mediated narrative discourse. Using the terminology of Russian formalists and a metaphor borrowed from cinema production, we introduce three models of increasing complexity. In the first model, the story logic (fabulist) creates a fabula which is adapted by the medium logic (director) to a screenplay then to the screen. In the second model, the story logic (screenwriter) generates a sjuzhet which is adapted to the screen by the medium logic (director). In the third model, the story and medium logics are co-authors of the mediated story. In all three cases, we propose a shared vocabulary of narrative discourse acts that can serve as an intermediate representation between story and medium ACM Subject Classification H.5.4 Hypertext/Hypermedia :Theory, J.5 Arts and humanities: Literature Keywords and phrases Narratology, Interactive drama, Media adaptation Digital Object Identifier /OASIcs.xxx.yyy.p 1 Introduction The domain of story generation covers a variety of computational techniques aiming at generating events that constitute a narrative. Within this paper, we will adopt a large acceptation of the term, that ranges from the generation of a text in a basic language, to the generation of pleasing aesthetic experiences, in text or visual media. We also include part of Interactive Digital Storytelling researches, when they make us of generative algorithms to adapt the story to the user s choices. The common feature of these techniques is that they require building an abstract computational model, that will be able to generate a variety of narrative events from a limited set of authored data. These models implement a story logic, as they formalize the occurrence of an event in the story according to various rules and algorithms. Depending on systems, this story logic can be based on the simulation of characters [7, 2], the simulation of reader s response [38, 3], the simulation of narrative acts [29, 32], etc. For this modeling task, there exist a vast set of narrative theories, in particular within the formalist approach (Propp, Bremond, Greimas). Ronfard and Szilas; licensed under Creative Commons License CC-BY Conference/workshop/symposium title on which this volume is based on. Editors: Billy Editor and Bill Editors; pp OpenAccess Series in Informatics Schloss Dagstuhl Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany

3 2 Narrative discourse At the same time, these stories aim to be expressed to an audience: the reader, the viewer, the listener, the user, depending on a specific medium. To represent a generated abstract story event within a certain medium requires another kind of theory. For text for example, the computational linguistics technique known as surface realization generates appropriate sentences from abstract clauses. Furthermore, an effective text generation system must also include a model of style and perform paraphrases to avoid monotonic language. Generating stories to different media such as 2-D or 3-D animation requires similar generative theories. Previous work in story generation has adopted, implicitly or explicitly, a simple communication model between story and medium: the pipeline model. Practically, it consists in taking the output of the story logic part, expressed as a series of actions or events, and feeding them to the medium logic. This approach has the advantage of simplicity but it raises two issues. First, the medium logic possesses limited information from the story logic and performs an uninformed flat representation of story events. Second, the story logic does not take into account the strengths and limitations of the medium. To resolve those issues, what is missing is a theory describing the interrelation between the story and the medium. In this paper, we lay the ground for such a theory by offering a review of previous work and proposing three models with increasing complexity. 2 Story and medium in narrative theories The relation between the narrative and the media is an old debate in narrative theories. Narratology as a discipline has emerged in the 1960s as a study of narrative structures that are independent of the medium [26]. For example, C. Bremond analyzes Propp s approach by stating that the structure of the story is independent of the techniques that carry it [6]. This independence is illustrated by the theory of G. Genette [12], which states that a narrative is composed of three distinct layers: the story consisting of events arranged in the temporal order within the fictional world; the discourse consisting of events re-organized in the temporal order in which they are presented to the audience; and the narration, which is the act of narrating a story and the concrete situation in which the story is conveyed through a physical media. This tripartite model finds its roots in Russian formalism, via the fabula/sjuzhet distinction. Note on that respect that one must remain careful with the terms coming from different narrative traditions and their translation between Russian, French and English. By adopting the working hypothesis that the story is independent from the other layers, theorists have produced several useful semi-formal models of the story [6, 13, 34]. Nevertheless, the hypothesis that narrative can be described totally independently of the media has been largely criticized in the theories of narrative: critics have been adamant that form does not separate from content [9]. According to many theorists, it is not possible to extract the story (logical layer) from one narrative (e.g. a book) to another media (e.g. film) without changing the story itself. Moreover, the relation between the story and the media can be inverted: this hypothesis suggests that the characteristics of a media determines, as a resource, which story can be told [17]. Herman proposes that between these two opposite positions (the thesis and the anti-thesis), a synthesis is possible [17]. It consists in considering that the media dependence of stories is a matter of degree. More precisely, some features of a narrative in a given media would translate well to another media, while other would not translate easily. As a corollary, some narratives, depending on which features of its media it uses, would more or less translate to another media. In this paper, we adopt Herman s position, that is we acknowledge the three layers from Genette s theory, but we will explore, from a computational point of view, what it means

4 Ronfard and Szilas 3 that some feature more or less translate across media. We will use the following concepts and terminology: Event: a fictional data that represents an elementary modification of the fictional world. Events include actions and happenings [24] Fabula: the set of events occuring in a narrative along with their temporal relations within the fictional world. Discourse: the set of events in the fabula along with their temporal relations within the narrative experience time. Mediated discourse (rendered story): the physical representation of the subset of events that are effectively displayed in a given media. Literary text, has been the implicit media in many theories of narrative, including Genette s theory. It has some specific features that constraint the type of narrative it can support: it is based on a complex written language; it is mono-modal (reading); it is sequential. In contrast, it is worth observing how cinema differs from literary text, and how it calls for a different account of the relation between fabula, discourse and mediated discourse [19]. The computer media, used in games and interactive storytelling works, albeit sharing properties with cinema provide yet another range of characteristics regarding the above-mentioned relation. In the rest of the paper, although 3-D virtual worlds is targetted in the short term, we will handle the above-defined narrative concepts in a broad accross media perspective. 3 Story and media in digital narratives For more than 30 years, research has been carried out in the fields of interactive fiction, interactive narrative, interactive drama. These researches consist of computer-generated stories, interactive or not, represented in various media: text, 2D graphics, 3-D. In this section, we want to explore how the resulting prototypes handle the relation between the fabula, the discourse and the rendered story. Because some of these systems have explicitly used narrative theories, we will also explore how they have interpreted the Genette s theory and other narrative models. Six cases will be discussed and contrasted, that represent (but do not pretend to cover) the state of the art in the domain. FearNot! [1] is an interactive drama prototype based on the simulation of autonomous agents. Agents use a complex cognitive and emotional architecture to generate actions dynamically, according to the current situation, that may be influenced by the user. In this generative system, the visualization in a 3-D game engine is clearly separated from the characters logic [2]. The actions themselves are completed with information regarding the way the action must be performed (e.g. facial animation). Because the outcome of the action depends on the physical configuration, this outcome is decided in the visualization engine, that sends this information back to the logical part of the system. In FearNot!, there is no explicit discourse layer: events in the fabula are simulated and then directly transmitted to the visualization engine. In addition, the visualization engine has no narrative information to indicate how to perform the action (importance of the action, of the character, level of suspense, etc.) and it executes it in a neutral manner. To sum up, FearNot! is based on a bipartite model with a strong independance between the two parts. The Mutiny [33] is a text-based interactive drama based on the IDtension system [32]. IDtension is inspired from strucuralist models of narrative. Narrative actions, inspired for example from T. Todorov analysis [34] are simulated and selected according to narrative criteria such as conflict or complexity. Once calculated in the narrative engine itself, fabula

5 4 Narrative discourse events and possible user choices are sent to another module that displays events as text and proposes choices via a specific menu system. Similarily to FearNot!, there is no separate discourse layer. However, the rendering layer does not render all events: it only renders events that involve the user, while other are logically executed but not displayed. As a result, the user may convey an information to Character A, and then receive a comment on this information from Character C, because in between, A talked to C. Therefore the visualization module carry out a typical function of the discourse layer, in Genette s theory. The resulting ellipses are interesting (the user must guess what happened) but not fully controlled by the system: the user might find unbelievable the fact that she does not perceive actions performed by characters just nearby. Nothing For Dinner [14] is also based on IDtension, but within a 3-D environment. It therefore fully demonstrates the advantage of media independence in terms of interoperability. The same story can be displayed in text or in 3-D [31]. Note however that the extension from text to 3-D was not straight-forward. Initially designed for text, IDtension would produce events sequentially. In 3-D, this situation was not acceptable, because the media calls for simultaneous actions occuring in parallel. It does not make sense if Character A waits for the user to have finished some action, before being able to talk to Character C. This illustrates the fact that a narrative engine, even if their creators claim it remains media-independant, may implicitly embed some medium-specific characteristics (in this case: non-simultaneity of action rendering). Nothing For Dinner also implements a noteworthy (lack of) management of visibility: The fact that the user sees or does not see the actions is never taken into account in the narrative engine. Therefore, sometimes the users would see what the characters do, sometimes they would not. This is left uncontrolled by the story logic and dependent on the current spatial configuration and user s movements. To summarize, the visualization engine in Nothing For Dinner does also play a role at the discourse level, but this role is not necessarily controlled narratively. Prom Week [18] is a facebook game, based on a large set of rules simulating the social relations between high school students. In terms of visualization, a cartoon representation of characters is used, which are involved in dialogs. While graphics themselves are basic and do not intend to convey any kind of narrative tension, it has been supplemented with a rich representation of internal features of characters: how they feel, how they relate to each other, their status, etc. This is a convenient way to avoid the Tale-Spin effect [37] that occurs when complex internal information behind characters actions are unfortunately hidden. In the case of Prom Week, the communication channel between the logic and media is large, which is made possible by the explicit display of internal information via numbers, icons, sliders, etc. This breaks with the tradition of many games and interactive drama that seek to minimize these information and promote a more immersive experience. Suspenser [8] is a module in a larger architecture that aims at generating suspenseful stories. This whole architecture is directly inspired from Genette s approach (adopting the russian terminology). Three main modules are considered: the Fabula Generator (producing the fabula), the Suspenser (transforming the fabula into the sjuzhet) and the Discourse Generator (transforming the sjuzjet into the media). The main innovation of this research is to explicitly tackle the transformation of the story into discourse (note however a shift in terminology, discourse here corresponds to the mediated discourse in our terms). Suspenser is able to automatically re-order events in the fabula to create a more suspenseful ordering of events. But it does without taking into account the specificities of the medium, which is handled by the Discourse Generator in a more traditional pipeline approach. Slant is a system for story generation [21] that integrates five components developed in

6 Ronfard and Szilas 5 three other systems developed independently. These components include: MEXICA that generates plots; Fig-S that generate variation of the plot by using metaphors; Verso that adds constraints regarding the genre; GRIOT-Gen that realizes metaphorical representations and Curveship-Gen that generates the text. The distinctive feature of Slant is that it goes beond the pipeline approach that characterizes all previous cases. Via a blackboard approach, the chain of processing is not always unidirectional. In particular, Verso can intervene after MEXICA by adding a new action to the plot and this action is in turn processed by MEXICA. Note however that these bottom-up processes are used for building the plot (corresponding to fabula in our terminology), not for building the medium-specific discourse. These six cases obviously do not cover all the field and many other cases would deserve a similar discussion, but this sampling is sufficient to formulate the following general observations: All systems more or less follow the general principle of independence between narrative layers. The separation between layers in not uniform across systems. In particular, for systems which are bipartite and not tripartie, the discourse layer may be dispatched in both modules. A strict separation between layers, that would consist in simulating a storyworld without taking into account the fact that this story is to be told and displayed, may not be possible without seriously sacrificing the core quality of the narrative. What is conveyed from the story layer to the discourse and/or media layer(s) varies among systems: from the mere ordered set of events to much more complex engine-specific structures, making interoperability between different systems difficult. In order to progress towards a computational model of story/media relation, we propose to focus our attention on the data that may circulate between story, discourse and media modules. In order to restrict our scope to the simplest case, even if it departs from a pure Genettian approach, we will only consider two modules. While this may appear limited in scope, this configuration is already sufficiently rich to open many new possibilitities in terms of narrative expressivity, as it will be detailed in the next sections. If a model of story/media relation is to reach a certain level of generality, the data that circulate between these two modules should be independent both from a specific narrative generation approach and from a specific media. As a result, the data should not refer to plans, speech acts or cases specific to a particular story logic and neither should they refer to verses, cameras or panels specific to a particular medium. The language used to communicate between story and medium should be neutral and yet expressive. We consider this language as a lingua franca, defined as "a language systematically (as opposed to occasionally, or casually) used to make communication possible between people not sharing a mother tongue, in particular when it is a third language, distinct from both mother tongues" (Wikipedia). In the next three sections, we will propose three successive specifications of a lingua franca between story and medium. They correspond to three options that may be adopted when designing a whole system for story generation, including the interactive storytelling case. These models are of increasing complexity, meaning that the first one is a special case of the second one which is a special case of the third one. Therefore the last model is the most sophisticated one. 4 The fabulist-director model As it has been described in the previous section, a common approach in most state-of-the-art systems is a pipeline model, where a story logic module creates fabula events and sends them

7 6 Narrative discourse to a media module. In case of a 3-D media, virtual actors play those events in real-time 3-D animation, and a cinematography module chooses camera viewpoints and displays them to the audience. Let us call this the fabulist and director model. The model has several advantages - it works and it is simple. Although we have already stressed the limitations of that model, it is important enough to be analyzed in more details. The core information that is needed to describe the fabula is the succession of events that happen in the fictional world. Therefore, events constitute the first elements in the lingua franca. Following the traditional disctinction in narratology [16, 24] an event may be either an action, in which case it involves an agent, or a happening, in which case no agent is causing the event. Events are usually described in a predicative form, where the predicate represent the class of event (expressed as a verb) and the parameters are role-value pairs such as (agent, character) in the case of an action. Other roles may take values from other elements: characters, objects, places, and events (in which case the events are nested). Therefore the lingua franca also include characters, objects and places, which are called existents [24]. Temporal relations between events can be described either implicitly or explicitly by providing a start time and a duration for each event. In a temporal media, the unfolding of a single event may take an unpredictable amount of time. Moreover, in some media, like interactive 3-D environments, the event may fail. Therefore, the medium needs to send back the information that the event is finished (eventfeedback), and the corresponding outcome, in terms of success or failure. Finally, the interactive case needs to be examined, regarding the lingua franca. There exists a large variety of interaction modes in the field of interactive storytelling but what is exchanged with the story module still consists of events. However, these events must be enriched with additional data. First, the story module needs to be able to send possible actions, the execution of which depending on the user s choice. Second, in return, the medium logic needs to inform the story logic that an action has been chosen. Therefore, in the lingua franca, the action is enriched with a status attribute, that can take three values: execution, user-possible, user-decided. Also, the lingua franca includes the case where the user is creating events that have not been proposed by the Fabulist. Therefore, a userevent element is introduced. Figure 1 summarizes the lingua franca related to the fabulist-director model. As already mentioned, this model is not new per se but illustrates one simple approach of story/medium relation, in which the story logic (fabulist) produces a raw description of actions that must be conveyed to the user by the medium logic (director). If the director uses the medium of text for example, it will use Natural Language Generation (NLG) techniques for surface realization of the predicate-based content, anaphora resolution, style adaptation, expressive paraphrasing and sentence re-ordering, etc. Similarly, in the medium of 3-D animation, the director must create character animation using either motion capture databases or procedural animation techniques, choose the staging and pacing of the animation and control the cinematography and the editing in real time [22]. Actions in the fabula may overlap or occur at the same time. Indeed, this can make the experience richer and more interesting. But it is then the role of the director to present such actions in an intelligible manner by controlling what the camera sees [25]. The choice of camera viewpoints is in fact very important to keep the audience engaged in the story and the camera can often be seen as a visual narrator [36]. The fabulist and director model has severe shortcomings. The director receives very little information from the fabulist to motivate directing choices: Is the event important or anecdotical? Which emotions are conveyed with this event? How this event relate to another

8 Ronfard and Szilas 7 Figure 1 Elements of the lingua franca for the fabulist-director model. event? Does this event relate to an existent? etc. Without answers to these questions, the director has no other option than to use standard, repetitive options. In the fabulist-director approach, the director s role is limited to showing events in the fabula in chronological order and with a neutral point of view. Even with such a minimal approach, it may be difficult or even possible to show all events to the viewer in the right order when they occur in parallel. In such cases, the director has no way of reporting what was actually shown to the viewer, which may cause the following steps in the story to become intelligible. 5 The screenwriter-director model To overcome the limitations of the previous model, it appears necessary to revise the role of the story logic. Instead of simply reporting fabula events, we now require that the story module communicates towards an audience, taking in charge (part of) the narrative discourse. By analogy with the case of film, this transforms the fabulist to a screenwriter. Indeed in traditional movie-making, it is a common practice to write an intermediate document - the screenplay - that represents events as they will appear in the movie (not the fabula) and from the point of view of the audience. Narratologically speaking, the screenplay is an interesting document, since it represents the narrative discourse in plain words, but with the temporal and spatial structure of a movie (scenes are indicated to represent spatial and temporal changes). However, our model departs from the film analogy by further imposing that the screenplay be medium-independent. Based on the above observations, we propose a different model of story-media relation where the story logic is not limited to creating fabula events, but also produces a narrative discourse as a series of discourse acts. In such discourse acts, the subject is the computer and the object is the audience. One outstanding issue in our proposal is how the screenplay should be represented, that is what are the new elements of the lingua franca. We of course cannot rely on natural language (as in a real screenplay) but propose a conceptual representation of

9 8 Narrative discourse narrative discourse acts. The main discourse act at work in a fictional discourse is CONVEY that simply consists of conveying information from the fabula to the player. Note that this discourse act is not part of the original speech act theory [28]. In fiction, the narrative discourse act of conveying a state or attitude or event to the player can best be compared to the speech act of pretending - which in Searle s theory is the core component of fiction [27]. The first difference with the fabulist-director model is that the screenwriter can now specifically choose the ordering of events. The following steps will further extend the model with typical discourse-related information. The first extension concerns states in the fabula. In the fabulist-director model, states in the fabula (object properties, mental states, etc.) are not communicated to the director, which is purely event-based. In the screenwriter-director model, the screenwriter may decide to convey, at a precise moment, a current state. For example, it can choose to convey the emotion of the character John: fear, just before this character attempts a dangerous action. Depending on the available medium, this information may be displayed differently. A director in the text medium may generate a sentence such as "John was terrified". A director in the 3-D animation medium may insert a close shot on John with the proper facial expression; or shake the camera or trigger a fearful music; etc. In complement to CON V EY discourse acts, narrative information can be added, in order to provide information on the manner to represent the action or happening. It includes the type of emotions that the event is expected to cause in the audience, the relation to other past of current events, the relation to characters and objects in the scene, the opening or closing of a subplot, etc. We therefore include a narrative indication field to the dicourse act, letting it open what kind of information may be included in this field. Another dimension of a discourse act is its relative importance of the event. Key actions, such as Barthes kernel functions [4], need to be represented with a specific focus. In film for example, the Hitchcock principle says that the size of an object that is currently in the frame should be related to its importance at that given moment [35]. Therefore, the model adds an importance field to the discourse act CONV EY. Regarding events ordering, major discourse-related processing such as flashbacks or flashforwards are processed, if any, within the screenwriter. However, the director may need some flexibility regarding the precise ordering of some overlaping events, both at the fabula and at the discourse level. For example, if Mary asks John for help for lifting a heavy box, the following event should occur immediately. But the decision of a third character Lucy to suddenly stand up to and walk to the fridge may occur now, or slightly later, without any significant change in the narrative. In some media, such as 3-D animation, such secondary action by Lucy may get in the way of the primary action involving John and Mary. This can be remedied by letting the screenwriter assign a priority indication to the discourse act CONV EY. Is the act of conveying a fabula event the only narrative discourse act at play? We found a couple of more subtle discourse acts in narrative across different media. In addition to the convey operator, it should be possible for the screenwriter to give indications that some fabula events should remain hidden from the player. We propose the operator NOT-CONVEY, meaning that a fabula event is taking place but hidden from the player until further notice. NOT CONV EY is not equivalent to an empty act, because the event does occur in the fabula, and the director must ensure that this event is not be perceived by the audience. For example, in a 3-D environment, if the director has received the information of not conveying the action of John lifting a box, it must ensure that the camera never displays the box and John lifting it. Interestingly, at the narrative level, this discourse act opens the way to decide

10 Ronfard and Szilas 9 Figure 2 Elements of the lingua franca for the screenwriter-director model

11 10 Narrative discourse later if the event happened or not, a strategy called late comitment [30], that allows for more flexibility in the narrative generation, especially in an interactive context. A more radical discourse act CONVEY-FALSE can be used to lead the audience to believe that some fabula event is happening, whereas it is in fact not the case. In some extreme cases, this may lead the director to lie to the audience, as in the 1995 movie Usual Suspect, by Bryan Singer, i.e. show events that did not take place in the fabula. In many cases, the same effect can be produced by providing only partial information from which the audience can draw false inferences. This creates an interesting twist when the audience then discovers what finally did "really" happen in the fabula. Such effect is subtle to render and, once again, is rendered differently by different media. In the 2011 movie The Artist, a modern silent movie by M. Hazanavicius, the main character is about to commit suicide, and a title with with the word bang occurs. While this seems to indicate that the character has shot himself, the next shot shows a car crashed against a tree! In this case, the effect is used for only a few seconds. In other examples, the wrong belief may last during the whole duration of the story. An extension of the NOT CONV EY discourse act concerns the characters rather than the audience. The screenwriter may wish to control which characters perceive the conveyed event. For example, it may specify that John is lifting a box but that Lucy does not perceives that. This applies for both CONV EY and CONV EY F ALSE. In the model, these discourse act are supplemented with a perceivedby field that contains one or more characters and their perception constraints (must perceive or must not perceive). Another very important discourse-related information is point of view (focalization). The screenwriter may decide that an event must be presented to the audience from a given perspective. This can be one of the participants in the event, or any other character known to perceive the event. In text generation, the Curveship system is able to change the point of view (who sees) as well as the narrator (who speaks) [20]. In 3-D environments, point of view is an important consideration for choosing camera angles [23]. Our model therefore includes an optional pointofview field that can contain characters or even objects. Finally, we introduces three additional narrative discourse acts to allow direct interaction with the audience: PROPOSE-EVENTS: The screenwriter proposes a list of possible events that the audience can choose from. Typically, in the case of interactive drama, it will include all actions from the user-controlled character. To each proposed event is attached one or more attributes that we de not specify at this level and that qualifies the choice. For example, the estimated suitability of playing this action at this moment may be provided, which could enable the Director to highlight most suitable choices. ENCOURAGE: Although interactivity is about giving choices to the user, it may be suitable in some context to influence the user towards a specific choice. Strategies of this kind have been suggested by researchers [38, 11]. Interestingly, there is a mirror effect between discourse level and the diegesis (fictional world), when an influence from a character serves the purpose of an influence from the enunciative instance. DISCOURAGE: It is the referse of the previous act: influencing the user so that she does not choose a given event. Similarily to the fabulist-director model, the user may also creates her own event. Figure 2 summarizes the lingua franca for the screewriter-director model. We do not believe that we have exhausted this configuration, yet, the model appears very rich compared to previous work. By no means do we recommand that a system implementing this lingua franca should be developed right away. The lingua franca should rather serve as an overview

12 Ronfard and Szilas 11 of the range of options that the screenwriter-director model offers, from which a system designer may pick whatever features appear relevant. 6 The co-authors model In the two previous models, we remained in a pipeline logic, where the flow of information remains uni-directional, from the story module to the media module. In this section, we propose a model that better accounts for the relation between story and media. The model builds upon the screenwriter and director model of the previous section, but adds one key element, which is bi-directional communication between the screenwriter and the director. The model considers that the story and the medium are two authors, collaborating to create a mediated narrative experience. To draw an analogy with film-making, it corresponds to the situation in which the screenplay is modified and re-written on the set, which is often the case in film production [10, 15]. The story logic is still in the position of generating discourse acts but the medium logic is now allowed to confirm, infirm and suggest narrative discourse acts as well. More precisely, rather than return success or failure, the medium can now send feedback in one of two forms: In case of success, the media module may execute a discourse act that is slightly different from the requested act. For example, it has conveyed the requested event but with an importance of 0.3 instead of 0.8. The media module produces the best effort to execute the request act but does not guarrantee that the event is represented exactly as requested. For that purpose, a new discourse act is introduced, CONFIRM, which include the details of how the the content has been effectively represented. Note that the media module has to make a decision wether the alternative discourse act is still acceptable or wether a failure return should be preferred (next case). In case of failure, and if it is possible to do so, the media module proposes an alternative discourse act that it could execute, that contains the same event or existent, but with different surrounding fields. For example, it may suggest to relax the constraint of perception (perceivedby) by authorizing that one character that was not allowed to perceive the action still perceives it. The corresponding discourse act is INFIRM, which contains (optionnally) a new set of values for the fields of importance, priority, perceivedby (or notperceivedby, perceivedfalseby) or pointofview. In the above case, the media module has been reactive, that is it has proposed an alternative act following an impossibility to perform the desired act. It can also be proactive, by suggesting not an event but a certains type of event. For example, in a 3-D environment, the director may suggest that the current spatial configuration of four characters into two groups of two would be suitable for conveying two simultaneous events with these two respective groups, one with a dialog, the other without; It could also suggest that an ominuous representation of a given character would be particularily suited at this moment (say with a low-angle shooting and a backlighting). Therefore, we introduce the discourse act of SUGGEST, containing the specification of fields characterizing an event. Figure 3 summarizes the three narrative discourse acts introduced above, which come in addition to those already present in Figure 2 (they replace the last two lines in Figure 2). We are not detailing here the process that should occur between the two "co-authors" that involves a sort of negotiation mechanism. Our focus remains on the lingua franca which now involves nine narrative discourse acts.

13 12 Narrative discourse Figure 3 Elements of the lingua franca for the story and media co-authors model. These elements, covering the media to story communication, come in addition to the elements described in the previous section that concerned the story to media communication (see Figure 2).

14 Ronfard and Szilas 13 7 Conclusion In this paper, we have proposed the first steps of a computational model of narrative that zooms out from the logic of story events to encompass the whole picture of narrative as an expressive artifact, embodied in a media. This has led us to focus on how the story logic and the medium logic need to converse, and to propose three models of what we have called the lingua franca between the story and the media. From a theoretical perspective, we had expected that the narratological debate regarding the narrative layers proposed by Genette [12] and others would have become simpler in the realm of computing. Indeed, while the notion of fabula in narratology is rather hypothetical, happening primarily in the head of the audience [5], the fabula in the computer program should be clearly defined and materialized, as a precise set of data handled by the algorithm for generation. However, both the observation of existing systems (Section 3) and our effort to build a model of fabula/discourse/media inter-relation show that separating programs in different layers is neither necessary nor sufficient to generate interesting and compelling stories in different media. Such considerations open the way to further narratological research on the complex relations between fabula, discourse and media in generative interactive drama. References 1 Ruth Aylett, Sandy Louchart, Joao Dias, Ana Paiva, and Marco Vala. Fearnot! - an experiment in emergent narrative. In Intelligent Virtual Agents. Springer, Ruth Aylett, Sandy Louchart, Joao Dias, Ana Paiva, Marco Vala, Sarah Woods, and Lynne Hall. Unscripted narrative for affectively driven characters. IEEE Journal of Graphics and Animation, 26(May/June):42 52, Paul Bailey. Searching for Storiness: Story-Generation from a Reader s Perspective. In Narrative Intelligence - Papers from the 1999 AAAI Fall Symposium - TR FS-99-01, pages , Menlo Park, CA, AAAI Press. 4 Roland Barthes. Introduction à l analyse structurale des récits. Communications, 8(1):1 27, David Bordwell. Narration in the Fiction Film. Univ of Wisconsin Press, Claude Bremond. Logique du récit. Seuil, Marc Cavazza, Fred Charles, and Steven J. Mead. Characters in Search of an author: AI-based Virtual Storytelling. In International Conference on Virtual Storytelling (ICVS 2001). LNCS 2197, Lecture Notes in Computer Science, pages Springer, Heidelberg, September Yun-Gyung Cheong and R Michael Young. Narrative Generation for Suspense: Modeling and Evaluation. In First Joint International Conference on Interactive Digital Storytelling (ICIDS). LNCS 5334, LNCS, pages , Berlin / Heidelberg, Springer. 9 K. Elliott. Literary film adaptation and the form/content dilemma. In Narrative Across Media. University of Nebraska Press, Syd Field. Screenplay: The Foundations of Screenwriting. Delta Trade Paperbacks, Rui Figueiredo and Ana Paiva. Affecting choices in interactive storytelling. In AAAI Fall Symposium in Computational Models of Narrative, November Gérard Genette. Figure III. Seuil, Paris, Algirdas Julien Greimas. Sémantique structurale. Presses universitaires de France, Paris, Nicolas Habonneau, Nicolas Szilas, Urs Richle, and Jean Dumas. 3d simulated interactive drama for teenagers coping with a traumatic brain injury in a parent. In ICIDS, 2012.

15 14 Narrative discourse 15 Matthew Paul Hawkins. Writing is Rewriting: Defining the Purpose of Drafts in Feature Screenplay Development in a Collaborative, Micro-budget Environment. Flinders University of South Australia, Screen and Media Studies, David Herman. Story Logic: Problems and Possibilities of Narrative. University of Nebraska Press, David Herman. Toward a transmedial narratology. In Narrative Across Media. University of Nebraska Press, Joshua McCoy, Mike Treanor, Ben Samuel, Michael Mateas, and Noah Wardrip-Fruin. Prom Week: social physics as gameplay. In FDG, Brian McFarlane. Novel To Film: An Introduction to the Theory of Adaptation. Clarendon Press, Nick Montfort. Curveship s automatic narrative variation. In Proceedings of the 6th International Conference on the Foundations of Digital Games (FDG 11), Nick Montfort, "Rafael P. Pérez", D. Fox Harrell, and Andrew Campana. Slant: A blackboard system to generate plot, figuration, and narrative discourse aspects of stories. In Proceedings of the Fourth International Conference on Computational Creativity, page , Sydney, Australia, jun Ken Perlin. Toward interactive narrative. In ICVS 05, Julie Porteous, Marc Cavazza, and Fred Charles. Narrative generation through characters point of view. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems: Volume 1 - Volume 1, AAMAS 10, pages , Richland, SC, International Foundation for Autonomous Agents and Multiagent Systems. 24 Gerald Prince. Dictionary of Narratology. University of Nebraska Press, Lincoln, NE, university edition, Nicholas Proferes. Film Directing Fundamentals: See Your Film Before Shooting. Focal Press, Marie-Laure Ryan. Introduction. In Marie-Laure Ryan, editor, Narrative Across Media. University of Nebraska Press, Lincoln and London, J. Searle. The logical status of fictional discourse. New Literary History, 6(2), John Searle. Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press, Nikita Sgouros. Dynamic Generation, Management and Resolution of Interactive Plots. Artificial Intelligence, 107(1):29 62, Ivo Swartjes and Mariët Theune. The Virtual Storyteller: story generation by simulation. In BNAIC, pages , Enschede, The Netherlands, University of Twente. 31 N. Szilas, T. Boggini, M. Axelradand P. Petta, and S. Rank. Specification of an open architecture for interactive storytelling. In ICIDS, volume LNCS, vol Springer, Nicolas Szilas. A computational model of an intelligent narrator for interactive narratives. Applied Artificial Intelligence, 21(8): , Nicolas Szilas. The mutiny: An interactive drama on idtension. In Proceedings of the 3rd International Conference on Digital Interactive Media in Entertainment and Arts, DIMEA 08, Tzvetan Todorov. Poetique de la prose. Seuil, Francois Truffaut and Helen G. Scott. Hitchcock. Simon & Schuster, Peter Verstraten. Film Narratology. University of Toronto Press, Noah Wardrip-Fruin. Expressive Processing: Digital Fictions, Computer Games, and SoftwareStudies. MIT Press, Peter Weyhrauch. Guiding Interactive Drama. PhD thesis, School of Computer Science, Carnegie Mellon Univ., 1997.